fix(agent/compaction): skipped truncating tiny tool outputs during pruning

- Added a 50-token minimum floor constant for pruning decisions.
- Allowed non-superseded, non-useless tool results below that floor to avoid truncation.

docs/compaction.md

@@ -151,6 +151,7 @@ Default prune policy:
 
 - Protect newest `40_000` tool-output tokens.
 - Require at least `20_000` total estimated savings.
+- Never blank a result below `50` tokens (`MIN_PRUNE_TOKENS`): the `[Output truncated - N tokens]` placeholder costs ~8 tokens, so pruning a sub-floor result would grow the context and churn the prompt cache for nothing. (Superseded and useless results keep their own rules — the useless collector already drops no-savings candidates; superseded reads prune for correctness regardless of size.)
 - Never prune `skill` tool results, `read` results of `skill://` paths, or reads of the active plan reference file (added via `AgentSession`'s plan protection).
 
 Pruned tool results are replaced with:

packages/agent/CHANGELOG.md

@@ -6,6 +6,10 @@
 - Added the `interruptible` tool field: when set, the agent loop may abort the tool mid-execution to deliver a queued steering message (honored only in `immediate` interrupt mode).
 - Added support for `gemini` and `gemma` as valid owned tool syntax values in environment configuration
 
+### Fixed
+
+- Fixed `pruneToolOutputs` blanking tiny tool results during overflow pruning: results below `50` tokens (`MIN_PRUNE_TOKENS`) are no longer replaced with the `[Output truncated - N tokens]` placeholder, which cost more tokens than the result itself and churned the prompt cache for zero savings.
+
 ## [15.13.2] - 2026-06-15
 
 ### Breaking Changes

packages/agent/src/compaction/pruning.ts

@@ -81,6 +81,16 @@ function createPrunedNotice(tokens: number): string {
 	return `[Output truncated - ${tokens} tokens]`;
 }
 
+/**
+ * Generic age-based pruning floor. Below this, blanking a result to
+ * `[Output truncated - N tokens]` recovers nothing — the placeholder itself
+ * costs ~8 tokens, so a sub-floor result grows the context (and churns the
+ * prompt cache) instead of shrinking it. Superseded/useless results keep their
+ * own rules: useless already drops no-savings candidates, superseded prunes for
+ * correctness regardless of size.
+ */
+const MIN_PRUNE_TOKENS = 50;
+
 function getToolResultMessage(entry: SessionEntry): ToolResultMessage | undefined {
 	if (entry.type !== "message") return undefined;
 	const message = entry.message as AgentMessage;
@@ -271,7 +281,8 @@ export function pruneToolOutputs(entries: SessionEntry[], config: PruneConfig =
 		// any age).
 		const superseded = supersededMessages?.has(message) ?? false;
 		const useless = uselessMessages?.has(message) ?? false;
-		if (!superseded && !useless && (accumulatedTokens < config.protectTokens || isProtected)) {
+		const tooSmall = tokens < MIN_PRUNE_TOKENS;
+		if (!superseded && !useless && (accumulatedTokens < config.protectTokens || isProtected || tooSmall)) {
 			accumulatedTokens += tokens;
 			continue;
 		}

packages/agent/test/supersede-prune.test.ts

@@ -503,3 +503,22 @@ describe("pruneToolOutputs — useless results", () => {
 		expect(resultText(result1)).toBe(NO_MATCH_TEXT);
 	});
 });
+
+describe("pruneToolOutputs — small-result floor", () => {
+	test("sub-floor results are left intact while a large neighbor is pruned", () => {
+		// "ok" is ~1 token: blanking it to `[Output truncated - 1 tokens]` would
+		// grow the context, so the floor must keep it. The large neighbor still prunes.
+		const [tinyCall, tinyResult] = readPair("src/tiny.ts", "ok", T0);
+		const [bigCall, bigResult] = readPair("src/big.ts", FILE_CONTENT, T0 + 1_000);
+		const entries: SessionEntry[] = [tinyCall, tinyResult, bigCall, bigResult];
+
+		// Protect window empty and zero savings threshold: only size keeps the tiny one.
+		const result = pruneToolOutputs(entries, { protectTokens: 0, minimumSavings: 0, protectedTools: [] });
+
+		expect(result.prunedCount).toBe(1);
+		expect(resultText(tinyResult)).toBe("ok");
+		expect(resultMessage(tinyResult).prunedAt).toBeUndefined();
+		expect(resultText(bigResult)).toMatch(/^\[Output truncated - \d+ tokens\]$/);
+		expect(resultMessage(bigResult).prunedAt).toBeDefined();
+	});
+});

packages/agent/test/tool-protection.test.ts

@@ -53,7 +53,7 @@ function readResult(toolCallId: string, text: string): SessionMessageEntry {
 describe("conditional tool-result protection", () => {
 	it("prunes regular read results but keeps skill:// reads", () => {
 		const skillResult = readResult("skill-read", "skill read output that must remain intact");
-		const fileResult = readResult("file-read", "file read output that can be pruned");
+		const fileResult = readResult("file-read", "file read output that can be pruned ".repeat(20));
 		const entries = [
 			assistantReadCall("skill-read", "skill://session-memory"),
 			skillResult,

packages/coding-agent/test/plan-mode/plan-protection.test.ts

@@ -123,7 +123,7 @@ describe("plan-read protection in compaction", () => {
 
 	it("prunes regular reads but keeps the plan read intact", () => {
 		const planResult = readResult("plan-read", "plan body that must remain intact");
-		const fileResult = readResult("file-read", "file body that can be pruned");
+		const fileResult = readResult("file-read", "file body that can be pruned ".repeat(20));
 		const entries = [
 			readCall("plan-read", "local://wp-migration.md"),
 			planResult,